Çalışanların eğitim ağının karmaşık ağ analizi

The advances in technology and data science affect many fields positively. One of these fields is education. Learning analytics have the potential to develop new ways of achieving excellence in teaching and learning. The companies try to use learning analytics techniques for their employees' education and aim to improve employee performance. The education data sets of Softtech employees are used in this study. Softtech is a software company in Turkey, and those data sets include different types of technical, non-technical, online, and offline education. All data sets are combined, and an employee network is created by connecting employees via education. In this study, complex network analysis, dynamic and static network analysis, link prediction, and machine learning techniques are applied with the aim of creating an education recommendation system. The study initially began with community detection in dynamic networks. During the application phase, it was concluded that due to the changes caused by the impact of the COVID-19 pandemic in the data, it was not suitable for community detection in dynamic networks. Therefore, it was decided to proceed with the link prediction method in the study. But the network is extensively examined as a dynamic network. In method section, all the methods used in the study are explained in detail. Firstly, information about how dynamic and static networks are generated is provided. Subsequently, information regarding the analysis of all micro, macro, meso, and time series required for the analysis of a complex network is shared. Descriptions of all the parameters used in the analyses are provided. After the information about network generation and analysis, a literature review on link prediction methods is shared. Afterwards, information regarding machine learning algorithms to be used for link prediction and performance evaluation metrics for these algorithms are shared. Finally, the method stage is completed by creating a template for machine learning-based link prediction for the education network of employees. When it comes to the application part, dynamic networks are firstly created. These dynamic networks are analyzed using the methods described in the method section, and the results are shared. There is data available for a period of 5 years, from January 2017 to December 2021. These data sets are separated on a monthly basis to create dynamic networks. At the end, a total of 59 monthly network files are obtained. Networks are generated from each network file. The connections between employees in relation to their education and the evolution of these connections are visualized on a monthly basis. Visuals from the year 2017 are shared to provide an idea. Sequentially, macro, micro, and meso analyses are conducted for each network. Based on these analysis results, the node counts, link counts, network structures, centrality measures, and community structures of each network are visualized. Information such as the node with the highest degree, the node with the highest betweenness centrality, and the nodes belonging to the same community can be accessed from these visualizations. Some of this information is shared to provide an idea. In the time series analysis section, each parameter is examined over time, and it is checked whether predictions can be made for the future. The impact of the COVID-19 pandemic is evident here. During the quarantine period, all employees work from home, resulting in an increase in the number of employees receiving education (node count) and the number of education sessions (link count) during the COVID-19 pandemic. Due to the lack of a regular distribution, successful results are not observed in terms of prediction. Additionally, the sudden decrease in the number of nodes (employees receiving education) in August 2018 indicates a period of mandatory collective leave. After analyzing and examining the data, the link prediction process is initiated. After explaining the raw data, the newly created education dictionary, and the newly generated static network structure, link prediction methods are applied in sequence. A total of 15 different link prediction methods are applied. Data until 2021 is used as training data, and data for 2021 is used as test data. These operations are performed using both training and test data. The output file contains 15 attributes corresponding to 15 methods. The output file also includes a label attribute indicating whether there is a connection between nodes in the network. Some of the connection prediction results are shared for illustrative purposes. Ultimately, two separate output files are generated, one for the training data and one for the test data, with 21 columns and 222778 rows each. The output files generated in the previous step are used as input files in the machine learning step. Prediction is being attempted using machine learning algorithms such as XGBoost, gradient boosting, random forest, logistic regression, support vector machines, and multilayer perceptron. Accuracy, precision, recall, and F1 score values are being examined to compare the accuracy of the models. The results of these values are shared in tables. Different preferences can be made for the model, but generally, support vector machine seems to be more preferred due to its consistency. All of its values are greater than 98%. The results indicate that the machine learning-based connection prediction method can be used in our study. It is observed that it is applicable to our data, and the results are usable. The subject of how to improve it is addressed. In the next stage, education information can also be recorded on the connections. The strength of the connection can be considered based on the number of shared education received. It is possible to advance and improve the study with different methods.